encode_thread() puts two MPVEncContexts (2*6516B here) on the stack and zeroes one of them in order to temporarily store the variables that get changed during encoding a macroblock (when there is more than one candidate type for a macroblock). This is wasteful and therefore this commit adds a small (328B here) structure to store exactly the fields that actually need to be backed up. Then one can extend MPVEncContext without fearing too use up to much stack.
Unified · Split
| @@ -2632,80 +2632,100 @@ static void encode_mb(MPVEncContext *const s, int motion_x, int motion_y) · 2631 unmodified lines | ||
| 2632 | 2632 | encode_mb_internal(s, motion_x, motion_y, 16, 16, 12, 0, 0, CHROMA_444); |
| 2633 | 2633 | } |
| 2634 | 2634 | |
| 2635 | −static inline void copy_context_before_encode(MPVEncContext *const d, | |
| 2636 | − const MPVEncContext *const s) | |
| 2637 | −{ | |
| 2638 | − int i; | |
| 2639 | − | |
| 2640 | − memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster than a loop? | |
| 2641 | − | |
| 2642 | − /* MPEG-1 */ | |
| 2643 | − d->c.mb_skip_run = s->c.mb_skip_run; | |
| 2644 | − for(i=0; i<3; i++) | |
| 2645 | − d->c.last_dc[i] = s->c.last_dc[i]; | |
| 2646 | − | |
| 2647 | − /* statistics */ | |
| 2648 | − d->mv_bits= s->mv_bits; | |
| 2649 | − d->i_tex_bits= s->i_tex_bits; | |
| 2650 | − d->p_tex_bits= s->p_tex_bits; | |
| 2651 | − d->i_count= s->i_count; | |
| 2652 | − d->misc_bits= s->misc_bits; | |
| 2653 | − d->last_bits= 0; | |
| 2654 | − | |
| 2655 | − d->c.mb_skipped = 0; | |
| 2656 | − d->c.qscale = s->c.qscale; | |
| 2657 | − d->dquant= s->dquant; | |
| 2658 | − | |
| 2659 | − d->esc3_level_length= s->esc3_level_length; | |
| 2635 | +typedef struct MBBackup { | |
| 2636 | + struct { | |
| 2637 | + int mv[2][4][2]; | |
| 2638 | + int last_mv[2][2][2]; | |
| 2639 | + int mv_type, mv_dir; | |
| 2640 | + int last_dc[3]; | |
| 2641 | + int mb_intra, mb_skipped, mb_skip_run; | |
| 2642 | + int qscale; | |
| 2643 | + int block_last_index[8]; | |
| 2644 | + int interlaced_dct; | |
| 2645 | + int16_t (*block)[64]; | |
| 2646 | + } c; | |
| 2647 | + int mv_bits, i_tex_bits, p_tex_bits, i_count, misc_bits, last_bits; | |
| 2648 | + int dquant; | |
| 2649 | + int esc3_level_length; | |
| 2650 | + PutBitContext pb, pb2, tex_pb; | |
| 2651 | +} MBBackup; | |
| 2652 | + | |
| 2653 | +#define COPY_CONTEXT(BEFORE, AFTER, DST_TYPE, SRC_TYPE) \ | |
| 2654 | +static inline void BEFORE ##_context_before_encode(DST_TYPE *const d, \ | |
| 2655 | + const SRC_TYPE *const s) \ | |
| 2656 | +{ \ | |
| 2657 | + /* FIXME is memcpy faster than a loop? */ \ | |
| 2658 | + memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); \ | |
| 2659 | + \ | |
| 2660 | + /* MPEG-1 */ \ | |
| 2661 | + d->c.mb_skip_run = s->c.mb_skip_run; \ | |
| 2662 | + for (int i = 0; i < 3; i++) \ | |
| 2663 | + d->c.last_dc[i] = s->c.last_dc[i]; \ | |
| 2664 | + \ | |
| 2665 | + /* statistics */ \ | |
| 2666 | + d->mv_bits = s->mv_bits; \ | |
| 2667 | + d->i_tex_bits = s->i_tex_bits; \ | |
| 2668 | + d->p_tex_bits = s->p_tex_bits; \ | |
| 2669 | + d->i_count = s->i_count; \ | |
| 2670 | + d->misc_bits = s->misc_bits; \ | |
| 2671 | + d->last_bits = 0; \ | |
| 2672 | + \ | |
| 2673 | + d->c.mb_skipped = 0; \ | |
| 2674 | + d->c.qscale = s->c.qscale; \ | |
| 2675 | + d->dquant = s->dquant; \ | |
| 2676 | + \ | |
| 2677 | + d->esc3_level_length = s->esc3_level_length; \ | |
| 2678 | +} \ | |
| 2679 | + \ | |
| 2680 | +static inline void AFTER ## _context_after_encode(DST_TYPE *const d, \ | |
| 2681 | + const SRC_TYPE *const s, \ | |
| 2682 | + int data_partitioning) \ | |
| 2683 | +{ \ | |
| 2684 | + /* FIXME is memcpy faster than a loop? */ \ | |
| 2685 | + memcpy(d->c.mv, s->c.mv, 2*4*2*sizeof(int)); \ | |
| 2686 | + memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); \ | |
| 2687 | + \ | |
| 2688 | + /* MPEG-1 */ \ | |
| 2689 | + d->c.mb_skip_run = s->c.mb_skip_run; \ | |
| 2690 | + for (int i = 0; i < 3; i++) \ | |
| 2691 | + d->c.last_dc[i] = s->c.last_dc[i]; \ | |
| 2692 | + \ | |
| 2693 | + /* statistics */ \ | |
| 2694 | + d->mv_bits = s->mv_bits; \ | |
| 2695 | + d->i_tex_bits = s->i_tex_bits; \ | |
| 2696 | + d->p_tex_bits = s->p_tex_bits; \ | |
| 2697 | + d->i_count = s->i_count; \ | |
| 2698 | + d->misc_bits = s->misc_bits; \ | |
| 2699 | + \ | |
| 2700 | + d->c.mb_intra = s->c.mb_intra; \ | |
| 2701 | + d->c.mb_skipped = s->c.mb_skipped; \ | |
| 2702 | + d->c.mv_type = s->c.mv_type; \ | |
| 2703 | + d->c.mv_dir = s->c.mv_dir; \ | |
| 2704 | + d->pb = s->pb; \ | |
| 2705 | + if (data_partitioning) { \ | |
| 2706 | + d->pb2 = s->pb2; \ | |
| 2707 | + d->tex_pb = s->tex_pb; \ | |
| 2708 | + } \ | |
| 2709 | + d->c.block = s->c.block; \ | |
| 2710 | + for (int i = 0; i < 8; i++) \ | |
| 2711 | + d->c.block_last_index[i] = s->c.block_last_index[i]; \ | |
| 2712 | + d->c.interlaced_dct = s->c.interlaced_dct; \ | |
| 2713 | + d->c.qscale = s->c.qscale; \ | |
| 2714 | + \ | |
| 2715 | + d->esc3_level_length = s->esc3_level_length; \ | |
| 2660 | 2716 | } |
| 2661 | 2717 | |
| 2662 | −static inline void copy_context_after_encode(MPVEncContext *const d, | |
| 2663 | − const MPVEncContext *const s, | |
| 2664 | − int data_partitioning) | |
| 2665 | −{ | |
| 2666 | − int i; | |
| 2667 | − | |
| 2668 | − memcpy(d->c.mv, s->c.mv, 2*4*2*sizeof(int)); | |
| 2669 | − memcpy(d->c.last_mv, s->c.last_mv, 2*2*2*sizeof(int)); //FIXME is memcpy faster than a loop? | |
| 2670 | − | |
| 2671 | − /* MPEG-1 */ | |
| 2672 | − d->c.mb_skip_run = s->c.mb_skip_run; | |
| 2673 | − for(i=0; i<3; i++) | |
| 2674 | − d->c.last_dc[i] = s->c.last_dc[i]; | |
| 2675 | − | |
| 2676 | − /* statistics */ | |
| 2677 | − d->mv_bits= s->mv_bits; | |
| 2678 | − d->i_tex_bits= s->i_tex_bits; | |
| 2679 | − d->p_tex_bits= s->p_tex_bits; | |
| 2680 | − d->i_count= s->i_count; | |
| 2681 | − d->misc_bits= s->misc_bits; | |
| 2682 | − | |
| 2683 | − d->c.mb_intra = s->c.mb_intra; | |
| 2684 | − d->c.mb_skipped = s->c.mb_skipped; | |
| 2685 | − d->c.mv_type = s->c.mv_type; | |
| 2686 | − d->c.mv_dir = s->c.mv_dir; | |
| 2687 | − d->pb= s->pb; | |
| 2688 | − if (data_partitioning) { | |
| 2689 | − d->pb2= s->pb2; | |
| 2690 | − d->tex_pb= s->tex_pb; | |
| 2691 | − } | |
| 2692 | − d->c.block = s->c.block; | |
| 2693 | − for(i=0; i<8; i++) | |
| 2694 | − d->c.block_last_index[i] = s->c.block_last_index[i]; | |
| 2695 | − d->c.interlaced_dct = s->c.interlaced_dct; | |
| 2696 | − d->c.qscale = s->c.qscale; | |
| 2697 | − | |
| 2698 | − d->esc3_level_length= s->esc3_level_length; | |
| 2699 | −} | |
| 2718 | +COPY_CONTEXT(backup, save, MBBackup, MPVEncContext) | |
| 2719 | +COPY_CONTEXT(reset, store, MPVEncContext, MBBackup) | |
| 2700 | 2720 | |
| 2701 | −static void encode_mb_hq(MPVEncContext *const s, MPVEncContext *const backup, MPVEncContext *const best, | |
| 2721 | +static void encode_mb_hq(MPVEncContext *const s, MBBackup *const backup, MBBackup *const best, | |
| 2702 | 2722 | PutBitContext pb[2], PutBitContext pb2[2], PutBitContext tex_pb[2], |
| 2703 | 2723 | int *dmin, int *next_block, int motion_x, int motion_y) |
| 2704 | 2724 | { |
| 2705 | 2725 | int score; |
| 2706 | 2726 | uint8_t *dest_backup[3]; |
| 2707 | 2727 | |
| 2708 | − copy_context_before_encode(s, backup); | |
| 2728 | + reset_context_before_encode(s, backup); | |
| 2709 | 2729 | |
| 2710 | 2730 | s->c.block = s->c.blocks[*next_block]; |
| 2711 | 2731 | s->pb = pb[*next_block]; |
| @@ -2745,7 +2765,7 @@ static void encode_mb_hq(MPVEncContext *const s, MPVEncContext *const backup, MP · 33 unmodified lines | ||
| 2745 | 2765 | *dmin= score; |
| 2746 | 2766 | *next_block^=1; |
| 2747 | 2767 | |
| 2748 | − copy_context_after_encode(best, s, s->c.data_partitioning); | |
| 2768 | + save_context_after_encode(best, s, s->c.data_partitioning); | |
| 2749 | 2769 | } |
| 2750 | 2770 | } |
| 2751 | 2771 | |
| @@ -2963,7 +2983,7 @@ static int encode_thread(AVCodecContext *c, void *arg){ · 211 unmodified lines | ||
| 2963 | 2983 | MPVEncContext *const s = *(void**)arg; |
| 2964 | 2984 | int chr_h = 16 >> s->c.chroma_y_shift; |
| 2965 | 2985 | int i; |
| 2966 | − MPVEncContext best_s = { 0 }, backup_s; | |
| 2986 | + MBBackup best_s = { 0 }, backup_s; | |
| 2967 | 2987 | uint8_t bit_buf[2][MAX_MB_BYTES]; |
| 2968 | 2988 | uint8_t bit_buf2[2][MAX_MB_BYTES]; |
| 2969 | 2989 | uint8_t bit_buf_tex[2][MAX_MB_BYTES]; |
| @@ -3164,7 +3184,7 @@ static int encode_thread(AVCodecContext *c, void *arg){ · 194 unmodified lines | ||
| 3164 | 3184 | int next_block=0; |
| 3165 | 3185 | int pb_bits_count, pb2_bits_count, tex_pb_bits_count; |
| 3166 | 3186 | |
| 3167 | − copy_context_before_encode(&backup_s, s); | |
| 3187 | + backup_context_before_encode(&backup_s, s); | |
| 3168 | 3188 | backup_s.pb= s->pb; |
| 3169 | 3189 | if (s->c.data_partitioning) { |
| 3170 | 3190 | backup_s.pb2= s->pb2; |
| @@ -3389,7 +3409,7 @@ static int encode_thread(AVCodecContext *c, void *arg){ · 218 unmodified lines | ||
| 3389 | 3409 | } |
| 3390 | 3410 | } |
| 3391 | 3411 | |
| 3392 | − copy_context_after_encode(s, &best_s, s->c.data_partitioning); | |
| 3412 | + store_context_after_encode(s, &best_s, s->c.data_partitioning); | |
| 3393 | 3413 | |
| 3394 | 3414 | pb_bits_count= put_bits_count(&s->pb); |
| 3395 | 3415 | flush_put_bits(&s->pb); |